Over the past 30 years, anti-fuse technology has attracted significant attention of many inventors, IC designers and manufacturers. An anti-fuse is a structure alterable to a conductive state, or in other words, an electronic device that changes state from not conducting to conducting. Equivalently, the binary states can be either one of high resistance and low resistance in response to electric stress, such as a programming voltage or current. There have been many attempts to develop and apply anti-fuses in microelectronic industry, but the most successful anti-fuse applications to date can be seen in FGPA devices manufactured by Actel and Quicklogic, and redundancy or option programming used in DRAM devices by Micron.
A summary of the progression of anti-fuse development follows as evidenced by issued United States patents.
Anti-fuse technology development started with U.S. Pat. No. 3,423,646, which disclosed a thin film formable diode PROM built as an array of horizontal and vertical conductors with a thin dielectric (aluminium oxide) between the conductors, at their crossings. Such NVM memory was programmed through perforation of the dielectric in some of the crossings. A formable diode would act as an open circuit until a voltage of sufficient magnitude and duration is applied to the crossing to cause forming of the aluminum oxide intermediate layer at which time device would act as a tunnelling diode.
U.S. Pat. No. 3,634,929 disclosed an inter-metal semiconductor anti-fuse array, the structure of the anti-fuse consisting of a thin dielectric capacitor (AlO2, SiO2 or Si3N4) utilizing two (Al) conductors located above and connected to the semiconductor diode.
A programmable dielectric ROM memory structure using a MOS capacitor and a MOS switching element was shown in U.S. Pat. No. 4,322,822 (McPherson). This cell was formed as a standard gate-oxide-over-substrate capacitor having a gate connected to a MOS transistor using a buried contact. In order to lower the oxide breakdown voltage, which needed to be smaller for the anti-fuse capacitor then for the MOS switch, a V-shaped grove in the capacitor area was proposed. Since the capacitor was formed between the poly gate and the grounded p-type substrate, the rupture voltage had to be applied to the capacitor through an access transistor. The Gate/Drain and Gate/Source edges of the access transistors were located at the second field oxide, much thicker then the gate oxide in the channel area, which greatly improved Gate/S-D breakdown voltage.
U.S. Pat. No. 4,507,757 (McElroy) proposed a method for lowering gate oxide breakdown voltage through avalanche junction breakdown. Although the original McElroy ideas evolved around using gated diodes to locally induce avalanche breakdown, which in turn lowered dielectric rupture voltage by enhanced electron tunnelling, he actually introduced or embodied other and perhaps more important elements to anti-fuse technology: (a) Dual gate oxide anti-fuse: access transistor gate oxide thicker then anti-fuse dielectric. McElroy's dual gate oxide process steps are: initial gate oxidation, etching areas for thinner gate oxide and subsequent gate oxidation. This procedure is now used in standard CMOS technologies for “I/O” and “1T” devices. (b) A “common-gate” (planar DRAM like) anti-fuse connection where access transistor connects to anti-fuse diffusion (Drain) node and all the anti-fuse gates are connected together. This is opposite to McPherson arrangement and results in much denser cell since the buried contact is eliminated. (c) Limiting resistor between common anti-fuse gate and external ground. (d) Two-terminal anti-fuse MOS device (a half transistor): McElroy concluded that only two terminals are needed in anti-fuse capacitor: D and G. The Source is not really needed for anti-fuse programming or operation and can be fully isolated from the active area. The bulk connection does not play any role either except for the avalanche breakdown. So the source role is limited to collecting carriers from the avalanche breakdown should the local substrate potential increase to forward bias the emitter of a parasitic n-p-n device formed by D, B and S.
It wasn't until 1985 when U.S. Pat. No. 4,543,594 (Mohsen) proposed an anti-fuse design suitable for redundancy repair. As such application requires much lower density than PROM, it was easier to supply external high voltage necessary to rupture the oxide without actually passing this voltage through the access transistors. Mohsen's anti-fuse structure consisted of a thin oxide (50-150A SiO2) polysilicon capacitor over a doped region. He believed that silicon from the substrate or silicon from the electrode where a polysilicon electrode is used melts into pin holes in the insulative layer to provide the conductor, and his test data showed that where the oxide layer is approximately 100 A thick and has an area between 10 to 500 um2, fusion occurred at a voltage of 12 to 16 volts. The current required to cause this fusion is less than 0.1 uA/um2 of capacitor area, and the resulting fused link has a resistance of approximately 0.5 to 2K ohms. A link, once fused, can handle currents of up to 100 milliamps at room temperature for approximately one second before it heals to an open fuse. Taking into account electron migration wear-out, the predicted wear-out lifetime of a link, once fused, is substantially greater than 3E8 hours.
The possibility of anti-fuse self-healing under current stress appeared to be the main roadblock for application of this technology in such areas like PROMs, PLDs and FPGAs, where constant fuse stress was required. The anti-fuse healing problem was resolved later by Mohsen and others at Actel in U.S. Pat. No. 4,823,181. Actel teaches the way to implement a reliable programmable low impedance anti-fuse element by using an ONO structure instead of silicon dioxide. Actel's method required an ohmic contact after dielectric rupture. This was achieved either by using heavily doped diffusion, or by putting an ONO dielectric between two metal electrodes (or silicide layers). The necessity of an Arsenic doped bottom diffusion electrode was revised later in U.S. Pat. No. 4,899,205, where it was allowed for either top-poly or bottom-diffusion to be highly doped.
U.S. Pat. No. 5,019,878 taught that if the drain is silicided, the application of a programming voltage in the range of ten to fifteen volts from the drain to the source reliably forms a melt filament across the channel region. A gate voltage may be applied to control the specific transistors to melt. IBM discovered similar effect by proposing a channel anti-fuse in U.S. Pat. No. 5,672,994. They discovered that with 0.5 um technology, the BVDSS for the nmos transistor is not only in the order of 6.5V, but once the S-D punch through occurs it creates permanent damage resulting in few kilo ohms leakage between the source and the drain.
U.S. Pat. Nos. 5,241,496 and 5,110,754 to Micron, disclosed a DRAM cell based anti-fuse (trench and stack). In 1996, Micron introduced a well-to-gate capacitor as an anti-fuse in U.S. Pat. No. 5,742,555. U.S. Pat. No. 6,087,707 proposed an N-Well coupled anti-fuse as a way to eliminate undercut defects associated with polysilicon etching. U.S. Patent Application No. 2002/0027,822 proposed a similar anti-fuse structure, but with n+regions removed to create an asymmetrical (“unbalanced”) high voltage access transistor using the N-well as a drain electrode.
U.S. Pat. No. 6,515,344 proposed a range of P+/N+ anti-fuse configurations, implemented using a minimum size gate between two opposite type diffusion regions.
NMOS anti-fuses have been built in an isolated P-well using a standard Deep N-Well process. Another variant of Deep N-Well based anti-fuses is disclosed in U.S. Pat. No. 6,611,040.
U.S. Patent Application Nos. 2002,0074,616 and 2004,0023,440 disclose other Deep N-Well anti-fuses. These anti-fuses consisted of a capacitor featuring direct tunnelling current rather then Fowler Nordheim current. These applications confirm that anti-fuse performance is generally improved for thinner gate oxide capacitors (approx 20 A, which is typical for transistors in 0.13 um process).
U.S. Pat. No. 6,580,145 disclosed a new version of a traditional anti-fuse structure utilizing dual gate oxides, with the thicker gate oxide being used for nmos (or pmos) access transistors and the thinner gate oxide for the capacitor. The N-Well (or P-Well) is used as a bottom plate of the anti-fuse capacitor.
The idea of creating a source drain short through the gate by separately breaking the S-G and D-G dielectric regions of the transistor is disclosed in U.S. Pat. No. 6,597,234.
U.S. Patent Application No. 2004,0004,269 disclosed an anti-fuse built from a MOS transistor having gate connected to the gate of a capacitor, degenerated by a thinner gate oxide and heavy doping under the channel through additional implantation (a diode). The rupture voltage is applied to a bottom plate of the capacitor.
In U.S. Pat. No. 6,667,902 (Peng) Peng attempts to improve a classic planar DRAM-like anti-fuse array by introducing “row program lines” which connect to the capacitors and run parallel to the word lines. If decoded, the row program lines can minimize exposure of access transistors to a high programming voltage, which would otherwise occur through already programmed cells. Peng and Fong further improve their array in U.S. Pat. No. 6,671,040 by adding a variable voltage controlling programming current, which allegedly controls the degree of gate oxide breakdown, allowing for multilevel or analog storage applications.
Most recently, U.S. Patent Application No. 2003/0202376 (Peng) shows a memory array using a single transistor structure. In the proposed memory cell, Peng eliminates the LDD diffusion from a regular NMOS transistor. A cross-point array structure is formed of horizontal active area (S/D) stripes crossing vertical poly gate stripes. Drain contacts are shared between neighbouring cells and connected to horizontal wordlines. Source regions are also shared and left floating. Peng assumes that if the LDD diffusion is omitted, the gate oxide breakdown location will be far enough from the drain area and a local N+ region will be created rather than D-G (drain-gate) short. If such a region was created, the programmed cells could be detected by positively biasing the gate and sensing the gate to drain current. In order to reduce the G-D or S-D (source-drain) short probability, Peng proposes increasing gate oxide thickness at the G-D and S_D edges through modification of a gate sidewall oxidation process. Peng's array requires that both source and drain regions be present in the memory cells, row wordlines coupled to transistor drain regions, and the column bitlines formed from transistor gates. Such an unusual connection must be very specific to Peng's programming and reading method, requiring a decoded high voltage (8V in 1.8V process) applied to all drain lines except for the one to be programmed. The decoded high voltage (8V) is applied to the gates of the column to be programmed, while the other gates are kept at 3.3V.
Although Peng achieves a cross-point memory architecture, his array requires CMOS process modifications (LDD elimination, thicker gate oxide at the edge) and has the following disadvantages: (a) All row decoders, column decoders and sense amplifiers must switch a wide range of voltages: 8V/3.3V/0V or 8V/1.8V/0V. (b) During a program operation, the 3.3V column drivers are effectively shorted to 8V row drivers or 0V drivers through programmed cells. This puts many limits on the array size, affects driver size and impacts reliability and effectiveness of programming. (c) Every program operation requires that all the array active areas (except for the programmed row) are biased at 8V. This leads to large N++ junction leakage current, and again limits array size. (d) The gate oxide breaking spot is assumed to be located far enough from the drain area so the punch through is not happening at 8V bias. At the same time, the transistor must operate correctly at 1.8V biasing—connecting to the channel area. This is not achievable without significant process modification. (e) Peng assumes that the gate oxide will not break on the source or drain edge if the LDD is not present. It is however known in the art that the S/D edges are the most likely locations for the oxide breakdown because of defects and electric field concentration around sharp edges.
Peng attempts to solve some of the high voltage switching problems in U.S. Patent Application No. 2003/0206467. The high blocking voltage on wordlines and bitlines is now replaced with “floating” wordlines and bitlines, and restrictions on the distance from the channel to the source and drain regions has been changed. Although floating wordlines and bitlines may ease problems with high voltage switching, they do not solve any of the above mentioned fundamental problems. Additionally they introduce severe coupling problems between the switched and the floating lines.
Today, anti-fuse developments concentrate around 3-dimensional thin film structures and special inter-metal materials. All these anti-fuse technologies require additional processing steps not available in standard CMOS process, prohibiting anti-fuse applications in typical VLSI and ASIC designs, where programmability could help overcome problems with ever shrinking device life cycles and constantly rising chip development costs. Therefore there is an apparent need in the industry for a reliable anti-fuse structures utilizing standard CMOS process.
Prior art anti-fuse cells and arrays either require special processing steps or suffer from high voltage exposure of MOS switching elements, leading to manufacturability and reliability problems. They are also limited to low density memory applications, with the exception of Peng's single transistor cell, which in turn has very doubtful manufacturability.
A significant issue with current non-volatile memories, such as Flash and OTP memories, is the speed at which data states of the memory cells can be sensed, which directly impacts overall performance of the memory. Performance of a memory, either embedded in a system or as a discrete memory device, can be the performance bottleneck for the system it is part of, relative to other processes executed by the system.
Non-volatile memories, such as Flash memories and OTP memories, use current sensing schemes as is well known in the art. These schemes are typically single ended, meaning that a sense amplifier circuit compares the current driven through one bitline which carries data of a memory cell connected to it, with a reference current. The reference current can be generated in a variety of ways, including synthesis by a reference voltage generator, or through a reference memory cell. The single bit digital output from a current sense amplifier represents the state of the bitline current relative to the reference current. In Flash memory, the current of a bitline will depend on the programmed threshold value of the memory cell. In an anti-fuse OTP memory, the current of a bitline will depend on the conductivity of the formed anti-fuse link.
Unfortunately, current sensing schemes are relatively slow. DRAM sensing on the other hand is much faster than current sensing schemes, since a voltage or charge is sensed on the bitlines. DRAM memories are organized in a folded bitline architecture, where pairs of bitlines are connected to their own bitline sense amplifier. Both bitlines (complementary) are precharged to some mid-point voltage level prior to a read operation, then a memory cell will either add to or remove charge from, one of the bitlines. Even a small voltage differential between the folded bitlines can be quickly detected by the bitline sense amplifier.
DRAM provides an optimal balance between high density and performance, which is why it is exclusively used for computer systems with ever-increasing demands for capacity and performance. In contrast, current anti-fuse OTP memories are relatively slow, but have useful non-volatile applications where DRAM is unsuitable or impractical to manufacture. Applications include onboard FLASH replacement, boot and processor code storage, PROM, EEPROM and EPROM replacement, MASK ROM replacement, and other applications where data must be securely retained in the absence of power. Unfortunately, even for such applications, the relatively slow performance of anti-fuse OTP memories can negatively impact the performance of the system that relies on the anti-fuse OTP memory, whether it is a set-top box, PDA, or cell phone.
It is, therefore, desirable to provide a simple and reliable, high density, anti-fuse array architecture suitable for implementation in standard CMOS technology, with high speed sensing performance.